qemu-block
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-block] [PATCH 4/9] mirror: efficiently zero out target


From: Eric Blake
Subject: Re: [Qemu-block] [PATCH 4/9] mirror: efficiently zero out target
Date: Tue, 14 Jun 2016 21:00:27 -0600
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:38.0) Gecko/20100101 Thunderbird/38.3.0

On 06/14/2016 09:25 AM, Denis V. Lunev wrote:
> With a bdrv_co_write_zeroes method on a target BDS zeroes will not be placed
> into the wire. Thus the target could be very efficiently zeroed out. This
> is should be done with the largest chunk possible.
> 
> This improves the performance of the live migration of the empty disk by
> 150 times if NBD supports write_zeroes.
> 
> Signed-off-by: Denis V. Lunev <address@hidden>
> Reviewed-by: Vladimir Sementsov-Ogievskiy<address@hidden>
> CC: Stefan Hajnoczi <address@hidden>
> CC: Fam Zheng <address@hidden>
> CC: Kevin Wolf <address@hidden>
> CC: Max Reitz <address@hidden>
> CC: Jeff Cody <address@hidden>
> CC: Eric Blake <address@hidden>
> ---
>  block/mirror.c | 32 +++++++++++++++++++++++++++++++-
>  1 file changed, 31 insertions(+), 1 deletion(-)
> 
> diff --git a/block/mirror.c b/block/mirror.c
> index c7b3639..c2f8773 100644
> --- a/block/mirror.c
> +++ b/block/mirror.c
> @@ -21,6 +21,7 @@
>  #include "qemu/ratelimit.h"
>  #include "qemu/bitmap.h"
>  
> +#define MIRROR_ZERO_CHUNK   (3u << (29 - BDRV_SECTOR_BITS))      /* 1.5 Gb */

Probably nicer to track this in bytes.  And do you really want a
hard-coded arbitrary limit, or is it better to live with
MIN_NON_ZERO(target_bs->bl.max_pwrite_zeroes, INT_MAX)?

> @@ -512,7 +513,8 @@ static int mirror_dirty_init(MirrorBlockJob *s)
>  
>      end = s->bdev_length / BDRV_SECTOR_SIZE;
>  
> -    if (base == NULL && !bdrv_has_zero_init(target_bs)) {
> +    if (base == NULL && !bdrv_has_zero_init(target_bs) &&
> +            target_bs->drv->bdrv_co_write_zeroes == NULL) {

Indentation is off, although if checkpatch.pl doesn't complain I guess
it doesn't matter that much.

Why should you care whether the target_bs->drv implements a callback?
Can't you just rely on the normal bdrv_*() functions to do the dirty
work of picking the most efficient implementation without you having to
bypass the block layer?  In fact, isn't that the whole goal of
bdrv_make_zero() - why not call that instead of reimplementing it?

Patch needs rebasing - we've redone this into bdrv_co_pwrite_zeroes and
a byte interface, since upstream commit c1499a5e.

>          bdrv_set_dirty_bitmap(s->dirty_bitmap, 0, end);
>          return 0;
>      }
> @@ -546,6 +548,34 @@ static int mirror_dirty_init(MirrorBlockJob *s)
>          }
>          sector_num += n;
>      }
> +
> +    if (base != NULL || bdrv_has_zero_init(target_bs)) {

You're now repeating the conditional that used to be 'bool
mark_all_dirty' (well, this is !mark_all_dirty); is it worth keeping the
simpler bool around?

> +        /* no need to zero out entire disk */
> +        return 0;
> +    }
> +
> +    for (sector_num = 0; sector_num < end; ) {
> +        int nb_sectors = MIN(MIRROR_ZERO_CHUNK, end - sector_num);

Why limit yourself to 1.5G? It's either too small for what you can
really do, or too large for what the device permits.  See my above
comment about MIN_NON_ZERO.

> +        int64_t now = qemu_clock_get_ns(QEMU_CLOCK_REALTIME);
> +
> +        if (now - last_pause_ns > SLICE_TIME) {
> +            last_pause_ns = now;
> +            block_job_sleep_ns(&s->common, QEMU_CLOCK_REALTIME, 0);
> +        }
> +
> +        if (block_job_is_cancelled(&s->common)) {
> +            return -EINTR;
> +        }
> +
> +        if (s->in_flight == MAX_IN_FLIGHT) {
> +            trace_mirror_yield(s, s->in_flight, s->buf_free_count, -1);
> +            mirror_wait_for_io(s);
> +            continue;
> +        }

Hmm - I guess your mirror yield points are why you couldn't just
directly use bdrv_make_zero(); but is that something where some code
refactoring can share more code rather than duplicating it?

> +
> +        mirror_do_zero_or_discard(s, sector_num, nb_sectors, false);
> +        sector_num += nb_sectors;
> +    }
>      return 0;
>  }
>  
> 

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org

Attachment: signature.asc
Description: OpenPGP digital signature


reply via email to

[Prev in Thread] Current Thread [Next in Thread]